Merge pull request #130 from moyu026/master

add mcp

Author: xing-yiren

Season2.step_into_llm/18.MCP/stdio_mcp.py

@@ -0,0 +1,109 @@
+#!/usr/local/bin/python
+# -*- coding: utf-8 -*-
+import pdb
+import logging
+import json
+import httpx
+from typing import Any
+from mcp.server.fastmcp import FastMCP
+
+proxies = {
+    "http://": httpx.Proxy("http://proxy-notebook.modelarts.com:8083"),
+    "https://": httpx.Proxy("http://proxy-notebook.modelarts.com:8083"),
+}
+
+# 初始化 MCP 服务器
+mcp = FastMCP("WeatherServer")
+ 
+# OpenWeather API 配置
+OPENWEATHER_API_BASE = "https://api.openweathermap.org/data/2.5/weather"
+API_KEY = " "# "请替换为你自己的 OpenWeather API Key"
+USER_AGENT = "weather-app/1.0"
+ 
+async def fetch_weather(city: str) -> dict[str, Any] | None:
+    """
+    从 OpenWeather API 获取天气信息。
+    :param city: 城市名称（需使用英文，如 Beijing）
+    :return: 天气数据字典；若出错返回包含 error 信息的字典
+    """
+    params = {
+        "q": city,
+        "appid": API_KEY,
+        "units": "metric",
+        "lang": "zh_cn"
+    }
+    headers = {"User-Agent": USER_AGENT}
+    print(OPENWEATHER_API_BASE, params, headers)
+    async with httpx.AsyncClient(proxies=proxies,timeout=30.0) as client:
+        try:
+            response = await client.get(OPENWEATHER_API_BASE, params=params, headers=headers)
+            response.raise_for_status()
+            return response.json()  # 返回字典类型
+        except httpx.HTTPStatusError as e:
+            return {"error": f"HTTP 错误: {e.response.status_code}"}
+        except Exception as e:
+            return {"error": f"请求失败: {str(e)}"}
+ 
+def format_weather(data: dict[str, Any] | str) -> str:
+    """
+    将天气数据格式化为易读文本。
+    :param data: 天气数据（可以是字典或 JSON 字符串）
+    :return: 格式化后的天气信息字符串
+    """
+    # 如果传入的是字符串，则先转换为字典
+    if isinstance(data, str):
+        try:
+            data = json.loads(data)
+        except Exception as e:
+            return f"无法解析天气数据: {e}"
+ 
+    # 如果数据中包含错误信息，直接返回错误提示
+    if "error" in data:
+        return f"⚠️ {data['error']}"
+ 
+    # 提取数据时做容错处理
+    city = data.get("name", "未知")
+    country = data.get("sys", {}).get("country", "未知")
+    temp = data.get("main", {}).get("temp", "N/A")
+    humidity = data.get("main", {}).get("humidity", "N/A")
+    wind_speed = data.get("wind", {}).get("speed", "N/A")
+    # weather 可能为空列表，因此用 [0] 前先提供默认字典
+    weather_list = data.get("weather", [{}])
+    description = weather_list[0].get("description", "未知")
+ 
+    return (
+        f"🌍 {city}, {country}\n"
+        f"🌡 温度: {temp}°C\n"
+        f"💧 湿度: {humidity}%\n"
+        f"🌬 风速: {wind_speed} m/s\n"
+        f"🌤 天气: {description}\n"
+    )
+ 
+@mcp.tool()
+async def query_weather(city: str) -> str:
+    """
+    输入指定城市的英文名称，返回今日天气查询结果。
+    :param city: 城市名称（需使用英文）
+    :return: 格式化后的天气信息
+    """
+    data = await fetch_weather(city)
+    print(data)
+    return format_weather(data)
+
+
+async def manual_get_weather():
+    await query_weather("beijing")
+
+import asyncio
+if __name__ == "__main__":
+    # 以标准 I/O 方式运行 MCP 服务器
+    mcp.run(transport='stdio')
+
+
+
+
+
+
+
+
+

Season2.step_into_llm/18.MCP/vllm_client.py

@@ -0,0 +1,151 @@
+import asyncio
+import json
+from openai import OpenAI
+from mcp.client.stdio import stdio_client
+from mcp import ClientSession, StdioServerParameters
+
+# 配置 vLLM 的 OpenAI 兼容 API
+OPENAI_API_KEY = "EMPTY"  # vLLM 不需要真实密钥
+OPENAI_API_BASE = "http://localhost:8000/v1"  # vLLM 的 OpenAI 兼容接口地址
+
+class MCPClientDemo:
+    def __init__(self, server_path: str):
+        self.server_path = server_path
+        # ✅ 使用 vLLM 的 OpenAI 兼容接口
+        self.llm = OpenAI(api_key=OPENAI_API_KEY, base_url=OPENAI_API_BASE)
+
+    async def run(self, user_query: str):
+        server_params = StdioServerParameters(command="python", args=[self.server_path])
+        async with stdio_client(server=server_params) as (read_stream, write_stream):
+            async with ClientSession(read_stream, write_stream) as session:
+                await session.initialize()
+
+                # 获取服务端注册的工具
+                tools = (await session.list_tools()).tools
+
+                # 转换为 OpenAI 格式的 tools
+                openai_tools = []
+                for tool in tools:
+                    openai_tools.append({
+                        "type": "function",
+                        "function": {
+                            "name": tool.name,
+                            "description": tool.description or "",
+                            "parameters": tool.inputSchema or {
+                                "type": "object",
+                                "properties": {
+                                    "city_name": {"type": "string", "description": "城市名称"}
+                                },
+                                "required": ["city_name"]
+                            }
+                        }
+                    })
+
+                # -------------------------------
+                # 方法 1: vLLM + MCP 工具调用
+                # -------------------------------
+                # 第一步：让模型判断是否需要调用工具
+                system_message = f"""
+                你是一个智能助手。你可以使用以下工具：
+                {json.dumps(openai_tools, ensure_ascii=False, indent=2)}
+
+                如果用户的问题需要调用工具，请回复 "TOOL_CALL: <工具名> <JSON参数>"。
+                如果不需要调用工具，请直接回答。
+                """
+
+                messages_for_tool_decision = [
+                    {"role": "system", "content": system_message},
+                    {"role": "user", "content": user_query}
+                ]
+                
+                # ✅ 使用 vLLM（通过 OpenAI 客户端）获取决策
+                try:
+                    decision_response = self.llm.chat.completions.create(
+                        model="/home/ma-user/work/Qwen2.5-1.5B-Instruct/",  # 模型名需与启动 vLLM 时一致
+                        messages=messages_for_tool_decision,
+                        max_tokens=512,
+                    )
+                    decision_text = decision_response.choices[0].message.content.strip()
+                except Exception as e:
+                    decision_text = ""
+                    print(f"调用 vLLM 失败: {e}")
+
+                result_with_tool = {"model_reply": "", "tool_called": None, "tool_result": None}
+
+                # 检查是否需要调用工具
+                if decision_text.startswith("TOOL_CALL:"):
+                    try:
+                        # 解析工具调用指令
+                        _, tool_name, args_json_str = decision_text.split(" ", 2)
+                        arguments = json.loads(args_json_str)
+
+                        # ✅ 通过 MCP 会话调用实际工具
+                        tool_result = await session.call_tool(tool_name, arguments)
+
+                        # 第二步：将工具结果返回给模型，生成最终回复
+                        messages_with_result = messages_for_tool_decision + [
+                            {"role": "assistant", "content": decision_text},
+                            {"role": "tool", "content": json.dumps(tool_result.model_dump(), ensure_ascii=False),
+                             "name": tool_name},
+                            {"role": "user", "content": "请根据以上工具调用结果，回答用户的问题。"}
+                        ]
+
+                        final_response = self.llm.chat.completions.create(
+                            model="/home/ma-user/work/Qwen2.5-1.5B-Instruct/",
+                            messages=messages_with_result,
+                            max_tokens=512,
+                        )
+                        result_with_tool["model_reply"] = final_response.choices[0].message.content
+                        result_with_tool["tool_called"] = tool_name
+                        result_with_tool["tool_arguments"] = arguments
+                        result_with_tool["tool_result"] = tool_result
+                    except Exception as e:
+                        result_with_tool["model_reply"] = f"工具调用解析错误: {e}。原始回复: {decision_text}"
+                else:
+                    result_with_tool["model_reply"] = decision_text
+
+                # -------------------------------
+                # 方法 2: 仅模型回复（无工具）
+                # -------------------------------
+                try:
+                    response_no_tool = self.llm.chat.completions.create(
+                        model="/home/ma-user/work/Qwen2.5-1.5B-Instruct/",
+                        messages=[{"role": "user", "content": user_query}],
+                        max_tokens=512,
+                    )
+                    message_no_tool = response_no_tool.choices[0].message
+                    result_no_tool = {
+                        "model_reply": message_no_tool.content
+                    }
+                except Exception as e:
+                    result_no_tool = {
+                        "model_reply": f"调用 vLLM 失败（无工具）: {e}"
+                    }
+
+                return {
+                    "user_query": user_query,
+                    "with_mcp_tool": result_with_tool,
+                    "without_tool": result_no_tool
+                }
+
+
+async def main():
+    client = MCPClientDemo(server_path="/home/ma-user/work/mcp/stdio_mcp.py")
+    result = await client.run("nanjing的天气怎么样")  
+
+    print(">>> 用户提问：", result["user_query"])
+    print("\n【使用 MCP 工具】")
+    print("模型回复：", result["with_mcp_tool"]["model_reply"])
+    if result["with_mcp_tool"]["tool_called"]:
+        print("调用工具：", result["with_mcp_tool"]["tool_called"])
+        print("工具参数：", result["with_mcp_tool"]["tool_arguments"])
+        print("工具结果：", result["with_mcp_tool"]["tool_result"])
+    else:
+        print("未调用任何工具")
+
+    print("\n【不使用工具】")
+    print("模型回复：", result["without_tool"]["model_reply"])
+
+
+if __name__ == "__main__":
+    asyncio.run(main())